more subtle optimisations

src/net/torvald/terrarum/audio/FFT.kt

@@ -12,6 +12,7 @@ private val IM0 = 1
 private val RE1 = -1
 private val IM1 = 0
 
+private val mulBuf = FloatArray(2)
 @JvmInline value class ComplexArray(val reim: FloatArray) {
 
 
@@ -21,12 +22,7 @@ private val IM1 = 0
         get() = reim.size / 2
 
     operator fun times(other: ComplexArray): ComplexArray {
-        val l = size
+        val out = FloatArray(size * 2) {
-//        val re = FloatArray(l) { res[2*it+RE] * other.res[it] - ims[it] * other.ims[it] }
-//        val im = FloatArray(l) { res[it] * other.ims[it] + ims[it] * other.res[it] }
-//        return ComplexArray(re, im)
-
-        val out = FloatArray(l * 2) {
             if (it % 2 == 0)
                 reim[it+RE0] * other.reim[it+RE0] - reim[it+IM0] * other.reim[it+IM0]
             else
@@ -37,6 +33,32 @@ private val IM1 = 0
         return ComplexArray(out)
     }
 
+    fun mult(other: ComplexArray, out: ComplexArray) {
+        for (it in 0 until size * 2) {
+            out.reim[it] =
+            if (it % 2 == 0)
+                reim[it+RE0] * other.reim[it+RE0] - reim[it+IM0] * other.reim[it+IM0]
+            else
+                reim[it+RE1] *other.reim[it+IM1] + reim[it+IM1] * other.reim[it+RE1]
+
+        }
+    }
+
+    // this is actually slower that having a separate array for mult results
+    /*fun inlineMult(other: ComplexArray) {
+        for (it in 0 until size * 2) {
+            mulBuf[it % 2] = if (it % 2 == 0)
+                reim[it+RE0] * other.reim[it+RE0] - reim[it+IM0] * other.reim[it+IM0]
+            else
+                reim[it+RE1] *other.reim[it+IM1] + reim[it+IM1] * other.reim[it+RE1]
+
+            if (it % 2 == 1) {
+                reim[it+RE1] = mulBuf[0]
+                reim[it+IM1] = mulBuf[1]
+            }
+        }
+    }*/
+
     fun getReal(): FloatArray {
         return FloatArray(size) { reim[it * 2] }
     }
@@ -120,6 +142,10 @@ object FFT: Disposable {
         return ComplexArray(signal)
     }
 
+    fun fft(signal0: ComplexArray) {
+        ffts[signal0.size]!!.complexForward(signal0.reim)
+    }
+
     // org.apache.commons.math3.transform.FastFouriesTransformer.java:404
     fun ifftAndGetReal(signal0: ComplexArray): FloatArray {
 //        val re = FloatArray(signal0.size) { signal0.reim[it * 2] }

src/net/torvald/terrarum/audio/dsp/Convolv.kt

@@ -109,6 +109,10 @@ class Convolv(ir: File, val gain: Float = 1f / 256f): TerrarumAudioFilter() {
     }
 
     private val realtime = (BLOCKSIZE / TerrarumAudioMixerTrack.SAMPLING_RATEF * 1000000000L)
+    private val fftIn = ComplexArray(FloatArray(fftLen * 2))
+    private val fftMult = ComplexArray(FloatArray(fftLen * 2))
+    private val fftOut = FloatArray(fftLen)
+
     /**
      * https://thewolfsound.com/fast-convolution-fft-based-overlap-add-overlap-save-partitioned/
      */
@@ -119,10 +123,16 @@ class Convolv(ir: File, val gain: Float = 1f / 256f): TerrarumAudioFilter() {
 
         for (ch in outbuf.indices) {
             push(inbuf[ch].applyGain(gain), this.inbuf[ch])
-            val inputFFT = FFT.fft(this.inbuf[ch])
+
-            val Y = inputFFT * convFFT[ch]
+            for (i in 0 until fftLen) {
-            val y = FFT.ifftAndGetReal(Y)
+                fftIn.reim[i * 2] = this.inbuf[ch][i]
-            System.arraycopy(y, fftLen - BLOCKSIZE, outbuf[ch], 0, BLOCKSIZE)
+                fftIn.reim[i * 2 + 1] = 0f
+            }
+
+            FFT.fft(fftIn)
+            fftIn.mult(convFFT[ch], fftMult)
+            FFT.ifftAndGetReal(fftMult, fftOut)
+            System.arraycopy(fftOut, fftLen - BLOCKSIZE, outbuf[ch], 0, BLOCKSIZE)
         }